Time series forecasting is essential across domains such as healthcare, energy, and climate modeling. While models like LSTMs, GRUs, Transformers, and State-Space Models (SSMs) have become widely used, selecting the optimal architecture remains unclear. We propose an automated framework that systematically designs hybrid architectures by combining LSTM, GRU, attention, and SSM modules. Our approach uses multi-objective optimization to explore combinations and orderings of blocks, yielding Pareto-optimal architectures that balance user-defined trade-offs among objectives. A preference function selects the most suitable model for a given application. Moreover, two sampling-based iterative procedures for Pareto-front exploration are introduced, which reduces the total training cost by nearly eightfold. Across four real-world benchmarks, our framework reveals that simple models excel in speed, while hybrid compositions dominate when balancing accuracy and complexity. Our findings challenge the notion of a universally superior neural architecture, emphasizing instead the value of data- and objective-driven design in time series forecasting.
Dynamic systems characterized by second-order nonlinear ordinary differential equations appear in many fields of physics and engineering. To solve these kinds of problems, time-consuming step-by-step numerical integration methods and convolution methods based on Volterra series in the time domain have been widely used. In contrast, this work develops an efficient generalized pole-residue method based on the Volterra series performed in the Laplace domain. The proposed method involves two steps: (1) the Volterra kernels are decoupled in terms of Laguerre polynomials, and (2) the partial response related to a single Laguerre polynomial is obtained analytically in terms of the pole-residue method. Compared to the traditional pole-residue method for a linear system, one of the novelties of the pole-residue method in this paper is how to deal with the higher-order poles and their corresponding coefficients. Because the proposed method derives an explicit, continuous response function of time, it is much more efficient than traditional numerical methods. Unlike the traditional Laplace domain method, the proposed method is applicable to arbitrary irregular excitations. Because the natural response, forced response and cross response are naturally obtained in the solution procedure, meaningful mathematical and physical insights are gained. In numerical studies, systems with a known equation of motion and an unknown equation of motion are investigated. For each system, regular excitations and complex irregular excitations with different parameters are studied. Numerical studies validate the good accuracy and high efficiency of the proposed method by comparing it with the fourth-order Runge--Kutta method.